sungkyunner/forceflow-bspline-knot-analysis
ForceFlow Adaptive B-spline Knot Analysis Training-free adaptive cubic B-spline fitting results on episode 0 of JokerESC/ForceFlow. No policy, SmolVLA model, or GPU was used. Question When action changes little but force/torque changes substantially, does fitting one vector-valued B-spline to [action, force, torque] allocate knots near the force event that are absent from an action-only fit? Both fits use one common knot vector per multichannel trajectory. For the… See the full description on the dataset page: https://huggingface.co/datasets/sungkyunner/forceflow-bspline-knot-analysis.
Add source-labeled reconstruction error experiment
Add full-episode reconstruction error threshold sweep
Run threshold sweeps for all selected events
Simplify headers and use behavior-oriented timeline labels
Use mid-scale behavior labels with preserved event transitions
Use broad motion and FT labels across videos and figures
Add full task video and broaden behavior labels
Add reconstruction RMSE threshold knot sweep
Define threshold sweep metrics in figures
Visualize action RMSE tolerance and fitting errors
Show all spline knot types in threshold sweep
Add original-camera video for clean-vase event 1388
Add legends and task-action tracks to every comparison panel
Replace regime codes with clean-vase task actions
Add FT-difference threshold sweep for clean-vase event 1388
Add joint action and FT spline to channel comparison
Show FT-guided comparison on action channel 3
Overlay complete action-only and FT-guided trajectories
Compare action-only and FT-guided knots on identical action channels
Add complete force-guided overall action view
Visualize all four action and FT regimes
Clarify selected-event knot metrics
Add force-guided action B-spline experiment
Split action-only and joint-fit action visualizations
Add trajectory geometry visualizations
Add ForceFlow adaptive B-spline knot analysis
initial commit